Robust Multimodal Sequence-Based Loop Closure Detection via

Transcription

Robust Multimodal Sequence-Based Loop Closure Detection via
An Inverse Optimal Control Approach for the
Transfer of Human Walking Motions in Constrained
Environment to Humanoid Robots
Debora Clever
Katja Mombaur
Optimization in Robotics & Biomechanics (ORB)
Interdisciplinary Center for Scientific Computing (IWR)
University of Heidelberg
[email protected]
Optimization in Robotics & Biomechanics (ORB)
Interdisciplinary Center for Scientific Computing (IWR)
University of Heidelberg
[email protected]
Abstract—In this paper we present an inverse optimal control
based transfer of motions from human experiments to humanoid
robots and apply it to walking in constrained environments. To
this end we introduce a 3D template model, which describes
motion on the basis of center of mass trajectory, foot trajectories,
upper body orientation and phase duration. Despite of its abstract
architecture with prismatic joints combined with damped series
elastic actuators instead of knees, the model (including dynamics
and constraints) is suitable to describe both, human and humanoid locomotion with appropriate parameters. We present and
apply an inverse optimal control approach to identify optimality
criteria based on human motion capture experiments. The identified optimal strategy is then transferred to the humanoid robot
for gait generation by solving an optimal control problem, which
takes into account the properties of the robot and differences in
the environment. The results of this approach are the center of
mass trajectory, the foot trajectories, the torso orientation, and
the single and double support phase durations for a sequence of
multiple steps allowing the humanoid robot to walk within a new
environment. We present one computational cycle (from motion
capture data to an optimized robot motion) for the example of
walking over irregular step stones with the aim to transfer the
motion to two very different humanoid robots (iCub Heidelberg01
and HRP-2 14). The transfer of these optimized robot motions
to the real robots by means of inverse kinematics is work in
progress and not part of this paper.
I. I NTRODUCTION
Humanoid gait generation in constrained environments is
a complex task which consists of a wide range of sub-tasks
reaching from scene recognition [11, 31] and state estimation
[27] over path planning [32] and pattern generation [17, 28]
and finally to the computation of a whole body motion based
on e.g. joint angles [21, 20]. To accomplish all those tasks, it
is desirable to observe humans themselves and prolongate the
gained knowledge to the robotic complements. As a specific
sub-task of the before mentioned challenges, the focus of
this paper is on the transfer of motion generation strategies
from humans to humanoids. Previous work on this topic
with focus on a direct transfer of joint angle trajectories
has been published by [6], transfer based on observations
in human gait on soft ground in [12]. Our approach goes
further in both directions. In contrast to [6] it allows for the
Fig. 1.
Starting from real motion capture experiments, locomotion is
transferred by means of identified optimality criteria for a dynamic template
model (top) to the humanoid robot models (middle). In a subsequent step (not
part of this paper) the motion is transferred to the real robots HeiCub and
HRP-2 using inverse kinematics (bottom).
generation of robot motions without performing human motion
experiments for this specific task. In contrast to [12] we use
a model based approach to obtain a mathematical description
of the observations made in human experiments. To this end
we base our work on the hypothesis that the human motor
system is optimal in its choice of motion [3, 4], use inverse
optimal control and a human multi-body model to identify
the underlying optimization criteria, adjust the optimal control
model to the robots constraints and finally use optimal control,
to compute a feasible robot motion based on the identified
criteria (see Figures 1,2).
To describe human and humanoid locomotion, in this work
we consider the same 3D dynamical template model, adjusted
to the specific embodiment of humans and robots by means
of parameters and constraints. The model consists of two legs
with prismatic joints combined with damped series elastic
actuators instead of knees, two point-masses as feet and a
reactive mass as upper body. Motion is described on basis
of center of mass trajectory, feet trajectories, upper body
orientation and phase duration [7].
Making use of template models for the identification of
optimality criteria is an interesting approach for robotic applications for the following reasons:
• The same model can be used for different walking
scenarios,
• The same model can be used for human gait analysis and
humanoid gait generation,
• A sequence of several steps can be considered,
• Computational results are directly usable for robot controllers if they are based on the same template model,
• It has potential to be used for robot control in real time.
In this paper we present one entire computational cycle,
starting from motion capture experiments and ending with
the generation of humanoid motions based on center of mass
trajectory, torso orientation, feet trajectories, and phase timing.
As a proof of concept we focus on one trial of constrained
walking and two humanoid robots. The first robot is the iCub
platform of Heidelberg University (HeiCub), a unique version
of iCub produced by iCub Facility Department of Fondazione
Istituto Italiano di Tecnologia (IIT) [15, 24]. It has 15 degrees
of freedom and consists of an upper body, a hip and two legs
with a total height of 0.97m and a total mass of 26kg. It does
neither have arms nor a head and is therefore quite similar to
the considered template model. The second robot is HRP-2 14
from Kawada, located at LAAS-CNRS [19, 18], a full body
bi-pedal robot with 30 degrees of freedom. It has a total height
of 1.54m and a total mass of 58kg.
As constrained environment we consider a step stone scenario, which is particularity interesting but also difficult,
because
• it combines several steps of different step length, direction
and duration for which common optimality criteria have
to be found,
• it includes constraints on the foot placement, and potential variation in ground height,
• it allows to consider significantly different scenarios for
optimality identification and for motion generation.
Note, that the presented approach is usable for a wide range
of different scenarios and can be used for the transfer to
basically any bi-pedal robot. In parallel, but not part of this
paper, we investigate how the identification of criteria transfers
between different subjects and how optimality criteria correlate
between different scenarios.
The paper is organized as follows: In Section two we
explain the general idea of inverse optimal control and the
solution strategy considered in this paper. In Section three we
data
human motion
inverse optimal
control
weights
objective
(for humans)
(with human dynamical model)
transfer rules
optimal control
robot motion
trajectories
(with robot dynamical model)
objective
(for robot,
usually robot specific)
Fig. 2. The role of inverse optimal control for motion transfer between
humans and humanoids based on optimality criteria.
present our 3D template model, which is suitable to mimic
human and humanoid motion at the same time, by adjusting
relevant parameters and constraints. In Section four we solve
the inverse optimal control problem, based on the presented
template model, a large set of optimality criteria for the
optimal control model and real motion capture data of a step
stone scenario. Finally, in Section five, the identified criteria
together with suitable transfer rules, mainly based on speed
issues and different bounds between human and humanoid,
are used to compute optimal motions for the two different
humanoid robot models each walking over a different step
stone scenario.
II. I NVERSE O PTIMAL C ONTROL FOR H UMAN G AIT
A NALYSIS AND H UMANOID G AIT G ENERATION
The overall goal of our research and the EU-project KoroiBot is to improve humanoid walking capabilities by learning
from humans. In this context inverse optimal control serves
as a tool to extract optimization criteria from human gait and
derive humanoid locomotion by minimizing these criteria with
respect to the robot’s architecture and bounds (see Figure 2).
A. Problem formulation
Based on the assumption that human gait is optimal [3, 4]
with respect to well defined but unknown criteria, we describe
human gait by a mathematical model in the form of an optimal
control problem restricted by a system of differential algebraic
equations (DAEs) and additional non-linear constraints on
control and state.
Given a reference motion Ẑ = (ẑi,tj )i=0,...,Nz −1, j=0,...,Nt̂ −1
describing Nz trajectories at discrete times (t̂0 , t̂1 , . . . , t̂Nt̂ −1 ),
we use inverse optimal control to determine an objective function Φ∗ (x, u, q, p, γ) such that the subset z(t) ⊂
(x(t), u(t), q) of the resulting optimal solution (x∗ , u∗ , q∗ )
approximates Ẑ (after scaling it with ω ∈ RNz ) as well as
possible.
Nt̂ −1
NX
z −1
X
1
ωi
(zi (tj ) − ẑi,tj )2 ,
min
(p,γ) Nz Nt̂
i=0
j=0
s.t. min Φ(x, u, q̃, p, γ)
(2)
(x,u,q̃)
s.t. ẋ = fk (x, u, q, p), k = 0, ... , ns −1
(eq)
rk (x, u, q, p) = 0,
(ineq)
rk
(x, u, q, p) ≥ 0,
ck (x(τ0 ), . . . , x(τns ), q, p) ≥ 0,
and
(0)
γ0 = γ0 .
(1)
(3)
(4)
(5)
(6)
The lower level (2)-(5) describes the optimal control problem of minimizing the objective function Φ(x, u, q, p) with
respect to the walking model (3)-(5), the upper level (1) is a
parameter estimation problem for the unknown weights γ ∈
RNγ := (γi )i=0,...,Nγ −1 and optionally some of the parameters
p ∈ RNp , such that the optimized motion z∗ ⊂ (x∗ , u∗ , q∗ )
is fitted to real motion capture data Ẑ. Equation (6) belongs
to the upper level problem and ensures uniqueness of the
parameter set γ by fixing γ0 to its initial value. The state
x : [t0 , tf ] → RNx and the control u : [t0 , tf ] → RNu are
vector functions in time, the control parameters q ∈ RNq and
the model parameters p ∈ RNp are vectors of real numbers.
The right hand side f := (f0 , . . . , fns −1 ) is phase-wise defined
(denoted by the subscript k) and allows for discontinuities
on phase boundaries. Switching condition between phases
are defined implicitly in the context of the constraints ck ,
which also cover other point-wise constraints like periodicity
requirements. As the right hand side, the path constraints rk
are defined phase-wise.
One important issue in the analysis of human gait are studies
on the duration of the different phases of the gait cycle - or
even more challenging: studies on optimality criteria leading
to realistic phase durations. To this end, the durations ∆τ
between the implicit defined switching points (τ0 , . . . , τns )
have to be included in the vector of control parameters q̃ :=
[q, ∆τ ]. Note, that there is a important difference between
the parameters q (and q̃, respectively) and p. Whereas q
is a control parameter of the lower problem, p is a model
parameter, and hence fixed for the lower problem. However, it
can be considered as an identification parameter on the upper
level.
Usually the reference trajectories Ẑ do neither include the
control nor do they define the entire state, but only a subset
of it. A special feature of the presented algorithm is that it
can handle arbitrary subsets z, since it does not build on
the assumption that for the reference motion all state and
control trajectories are explicitly given. It can even handle
the special case, if Ẑ does not correspond to a subset z(t) ⊂
(x(t), u(t), q) but to a general function M(x(t), u(t), q).
In our research we define the lower objective function of
the optimal control model
PNγ −1(2)-(5) as a linear combination
Φ(x, u, q, p, γ) :=
i=0 γi φi (x, u, q, p), of physically
meaningful sub-criteria φi , i = 0, . . . , Nγ − 1 which can describe common criteria known from biomechanics or robotics
or new hypotheses based on human experiments. There are two
major advantages to use this kind of objective function: first,
using physically meaningful sub-criteria, the results of gait
analysis are physically interpretable and second, it implies the
possibility to derive transfer rules from humans to humanoids
which can take into account dynamic and kinematic differences between the two embodiments.
B. Solution strategy
To solve the inverse optimal control problem (1) - (6) there
are to two main types of solution strategies. In the first strategy
the bi-level structure of the problem is kept and exploited [25,
26, 29]. In the second one the bi-level problem is replaced by
a one-level problem, substituting the lower level problem by
its optimality conditions [8, 9, 14, 5, 13, 2]. In our research
we consider and also compare both approaches. In this paper
we focus on the first approach which has shown to perform
well for real-life optimal control problems and allows for an
easy implementation of new problems. An one-level approach
for similar studies, but with an easier template model walking
on level ground, has been considered in complementary work.
As both, the upper and the lower level of the bi-level optimal
control problem are optimization problems themselves, we
need suitable solution strategies. Whereas the upper level is a
parameter identification problem with constraints of unusual
form, the lower level problem is restricted by a hybrid differential equation with additional linear and nonlinear constraints, where known structures can be exploited. Therefore
on the upper level we rely on a derivative free optimization
method, in this paper the quadratic approximation based
method BOBYQA [30], implemented in the optimization
library NLOPT [16]. For the lower level, we use a derivative
based approach which relies on a state parametrization by
multiple shooting and a structure exploiting SQP algorithm to
solve the resulting discrete nonlinear programs (NLPs). Both
algorithms have been combined in an efficient inverse optimal
control framework that allows the user
• to define an optimal control model, which consists of an
objective function defined by a linear combination of optimization criteria, a dynamical model given by a hybrid
differential equation, and additional linear and nonlinear
constraints (also specifying implicit phase changes),
• to define the sub-set of optimization parameters to be
taken into account on the upper level, which usually
consists of the objective weights γ and a subset of model
parameters p,
• to define the sub-set of quantities which are the ones
that are fitted to motion capture data or functions thereof.
These quantities are usually a subset of the state x, and
the subset ∆τ ⊂ q, defining the duration of the different
phases.
III. L OCOMOTION AS O PTIMAL C ONTROL M ODEL
As mentioned above, in our work we model locomotion
as an optimal control problem, relying on the hypothesis that
the human motor control system is optimal in its choice of
motions. For our framework, such an optimal control model
consists of three main parts. First, the dynamic multi-body
model, which we usually model on bases of Newtons law,
reformulating the differential algebraic equations of index
three as a system index one. Second, the gait cycle as
a sequence of changing contact sets and additional linear
and non-linear constraints. Different dynamical behavior as
a consequence of varying contact sets during the gait cycle is
defined by different phases, resulting in a hybrid differential
equation. Phase durations are defined implicitly by characteristic state and control dependent switching events. Furthermore,
we explicitly allow for discontinuities in the state functions
to model the impact of foot touch down. Third, a twice
continuously differentiable objective function. As stated above,
in the context of inverse optimal control, we formulate the
objective as a linear combination of different criteria. Note,
that the criteria themselves can be highly non-linear.
A. The dynamical walking model and the step stones
(TD)
(TD)
(TD)
(TD)
(SS)leg 1 (DS) (SS)leg 2 (DS) (SS)leg 1 (DS) (SS)leg 2 (DS)
Fig. 4.
discontinuities
continuous
phases
Sequence of walking phases.
(TD), and double support phase (DS), see Figure 4. In our
computations, we consider a sequence of two full strides,
where each stride contains two steps with alternating legs.
During the different phases of one single step, we formulate
the following right hand sides and additional constraints:
• Single support phase (SS) on leg i, with swing foot j
RHSmi = 0,
Fig. 3.
(11)
uxj
uyj
Fj
vj +
nj − mj g, (12)
RHSmj = − (qM − qmj ) +
lj
lj
lj
lj ≤ l0 + usj , zmj ≥ h(xmj , ymj ),
(13)
3D SEA-walker
We consider the three dimensional dynamical template
model (3D SEA-walker), which we have first introduced in
[7], and slightly tailor it to the arising needs. It consists of
an upper body (torso) and two legs, which are attached to the
center of mass of the torso. Both legs have a compliant knee,
modeled by a damped series elastic actuator, and a point foot
with appropriate mass.
Denoting the absolute positions of the center of mass of the
torso and feet positions by
 




xM
xm1
xm2
qM :=  yM  , qm1 :=  ym1  , qm2 :=  ym2  ,
zM
zm1
zm2
and defining the torso mass M , the feet masses m1 and m2 , the
leg length of the i-th leg li , and the inertia tensor of the torso
Θ, the dynamics of the motion is defined by the following
equations:
F2
F1
(qM − qm1 ) +
(qM − qm2 ) − M g
l1
l2
= RHSm1 , m2 q̈m2 = RHSm2
M q̈M =
m1 q̈m1
Θα̈ = uτ 1 + uτ 2 .
(7)
(8)
(9)
We set
 
0
αx
ux1
ux2
α :=
, uτ 1 :=
, uτ 2 :=
, g := 0 ,
αy
uy1
uy2
g
with gravitational acceleration g. Including a series elastic
actuator (SEA) and a damper, the leg forces are given as
Fi = −k(li − l0 − usi ) − b · l˙i ,
(10)
where k is the spring constant, b the damping constant, li the
leg length, l0 the rest length, us = (us1 , us2 )T the control
of the SEA, uτ = (uτ 1 , uτ 2 )T the hip torque, and α the
torso orientation. Each step is modeled as a sequence of single
support phase (SS), instantaneous touch down discontinuity
• Discontinuity at touch down (TD), leg j
q̇mj = 0,
(xmj , ymj ) ∈ Cstepstones ,
zmj = h(xmj , ymj )
(14)
(15)
• Transition conditions (SS, leg i) → (DS, leg j in front)
zmj = h(xmj , ymj ),
żmj ≤ 0,
(16)
• Double support phase (DS, leg j in front)
RHSmi = 0,
RHSmj = 0,
li ≤ l0 + usi ,
(17)
lj ≤ l0 + usj ,
(18)
• Transition conditions (DS, leg j in front) → (SS, leg j)
grf z,i = 0,
l˙i ≥ c > 0,
lj ≤ l0 + usj ,
(19)
where grf z,i is the vertical ground reaction force of leg i.
Furthermore we fix (xM , yM ) at initial time t = 0 and
final time T to a desired value, to allow to use the same
model for later gait generation. The subsequent step is defined
analogously to (11)-(19) by switching the roles of leg i and j.
Different from the version of the SEA-walker in [7], as
upper body we consider a regular cone, with center of mass
2
2
2
(CoM) at h4t and inertia tensor Θ = M
80 diag(12rt +3ht , 12rT +
3h2t , 24h2t ) (cone radius rt , cone height ht ). Furthermore we
consider the torso orientation α = (αx , αy ) around the global
x and y axis, in analogy to the absolute positions of CoM and
feet. The vector vi lies in th y-z plane and is perpendicular
to the vector from qmi to qM , the vector ni is lies in the
x-z plane and is also perpendicular to the vector from qmi to
qM . The torque uxi implies a linear force in direction of vi
on foot i, the torque uyi implies a linear force in direction of
ni on foot i. The function h : R2 → R defines the height of
the uneven ground, the set Cstepstones the feasible region to be
stepped on.
We parameterize the ns step stones by Ns = 4 · 3 · (ns )
parameters pi , which correspond to the markers placed on
the corners of each step stone during the experiments. Therefore, for each considered motion, the step stone model is
automatically adjusted to position, orientation and dimension
of the real blocks. Feasible (x, y)-regions are described by
the four lines, passing through the (x, y) coordinates of two
adjacent markers, with its normal pointing inwards. A foot
placement is defined to be feasible, if the distance between
the (x, y)-position of the foot and all four lines is greater
than a safety margin and if the position (x, y, z) fulfills the
condition h(x, y) = z, with function h describing the height
of the ground or the step stones at a point (x, y).
B. Optimality criteria
In this work, we propose a set of twelve different optimality
criteria φi , i = 0, . . . , 11, depending on state x, control
u, and one optimality parameter pDSSS which defines the
optimal ratio of double and single support. Note, that all
objective functions of Lagrange type are additionally divided
by the corresponding phase duration ∆τk to avoid an implicit
minimization of phase times. As described above, the objective
function to be identified is defined as follows
Φ(x, u, pDSSS ) =
11
X
γi φi (·),
(20)
i=0
minimization of the SEA actuation in the swing foot
Z
Z
φ1 (u) = min
u2s2 dt + min
u2s1 dt,
SS1
(21)
(22)
SS2
minimization of hip torque of the swing foot
Z
Z
φ2 (u) = min
kuτ 2 k2 dt +min
kuτ 1 k2 dt,
(23)
SS2
minimization of angular momentum in x-direction
Z T
φ3 (x) = min
(Θx · α̇x )2 dt,
(24)
0
minimization of angular momentum in y-direction
Z T
φ4 (x) = min
(Θy · α̇y )2 dt,
(25)
0
minimization of vertical center of mass oscillations
Z T
2
φ5 (x) = min
żM
dt,
(26)
0
minimization of absolute swing foot velocity
Z
Z
φ6 (x) = min
kq̇m2 k2 dt + min
kq̇m1 k2 dt, (27)
SS1
(28)
minimization of periodicity gap in center of mass velocities
φ8 (x) = min kq̇M (0) − q̇M (T )k2
(29)
minimization/maximization of overall single support duration
X
φ9 (x) = ±
qt,ss1 + qt,ss2 ,
(30)
all steps
minimization of absolute swing foot velocity at touch down
X
φ10 (x) =
min kq̇m1 (ttd1 )k2 + kq̇m2 (ttd2 )k2 ,
(31)
all steps
tracking of ratio between sub-sequent double and single support phase to constant but unknown value pDSSS
2
X qt,ds
(32)
φ11 (x) =
qt,ss − pDSSS .
IV. O PTIMALITY C RITERIA I DENTIFICATION
DS
SS1
all steps
+ k(xm2 (ttd2 ), ym2 (ttd2 ))T − ξ(ttd2 )k ,
s
l0
(ẋM (t), ẏM (t))T
ξ(t) = (xM (t), yM (t))T +
g
all steps
with the optimality functions of Lagrange type:
minimization of the SEA actuation in the stance foot
Z
φ0 (u) = min
u2s1 dt
ZSS1
Z
2
+ min
us2 dt + min
u2s1 + u2s2 dt,
SS2
and the optimality functions of Mayer type
minimization of planar distance between foot position at touch
down and capture point
X
φ7 (x) =
min k(xm1 (ttd1 ), ym1 (ttd1 ))T − ξ(ttd1 )k
SS2
For the identification of optimality criteria, or to be more
precise, of the optimality weights γ and optimality criteria
related parameters p, we rely on a least square fit of model
CoM, feet trajectories, torso orientation, and phase durations
to the corresponding reference trajectories and reference times.
In accordance to common notation in optimal control of
ODEs and DAEs and with respect to Section II, in the
following we refer to the control as u := (us , uτ ) and to the
state as x := (qM , qm1 , qm2 , α, q̇M , q̇m1 , q̇m2 , α̇). However,
x as a subscript denotes the x-axis of the basis frame.
A. Model parameters and reference data
As in the first part of the presented motion transfer, the
3D-SEA walker is meant to mimic a human being, the
model parameters have to be set as close as possible to the
corresponding values of the subject and the environment of
the motion capture experiment.
Environment and subject specific data (such as block size
and location, leg length, torso height, etc.) can directly be
measured. Feet and upper body masses are approximated based
on de Leva [10] and the absolute weight of the subject. The
width of the cone is set to rt := 21 ht . Spring constant k and
damping factor b are chosen by use of optimal control, such
that for a minimized amount of actuation the model mimics the
main characteristics of human walking. All relevant parameters
are given in Table I.
We use motion capture and subject specific data, jointly
recorded with M. Giese et al. at Eberhard Karls University,
TABLE I
WALKING MODEL PARAMETERS AND BOUNDS FOR HUMAN SUBJECT
human subject
body weight
est. upper body weight
foot weight
leg rest length
torso height
torso radius
spring stiffness
damping constant
bound ss duration (l,u)
bound ds duration (l,u)
safety margin on block
(front, left, back, right)
60kg
30kg
0.3 kg
0.965m
0.7m
0.35m
3300kg/s2
660 Ms/m
0.4s, 0.6s
0.2s, 0.5s
(0.02, 0.02, 0.02, 0.02)m
Tübingen, and published in the open source KoroiBot motion
database, set up by Karlsruhe Institute of Technology [23, 1].
To convert the recorded motion to model related reference
data we extract a mean hip point qM,ref = 14 (LASI ∗RASI ∗
LHIP ∗ RHIP ), mean foot points (xm,ref , ym,ref ) =
1
2 (T OE + HEEL), zm,ref = min(T OE, HEEL), and
the angles αx and αy given by the orientation of a
line through the mean hip point qM,ref and the marker
C7. LASI, RASI, LHIP, RHIP, T OE, HEEL and C7 are
marker labels and denote markers for which three dimensional
position trajectories are recorded. For a full description of the
used marker set (and its labeling) we refer to the marker set,
introduced in [22].
The reference motion starts, when the rear leg is about to
lift off from the floor while the front leg is already on the first
tread. The motion ends after four full steps, when the front
leg is on the last step stone and the rear leg is about to take
off again.
motion of 0.04m and an average difference in phase times of
0.11s. After 121 iterations on the upper level the algorithm
converges with
γ̃ ∗≈ {1.0, 1.3, 0.8, 0.92, 0.93, 0.91, 0.91, 1.35, 1.28, 0.51, 0.98, 0.1, 1.61}.
For the identified weights the average quadratic deviation of
the optimized motion can be reduced to 0.03m and 0.06s.
A comparison between a representative subset of the computed quantities to their reference counterpart is presented in
Figure 5, a comparison of phase durations in Figure 6.
Fig. 5. CoM trajectory and left foot trajectory resulting from identified
objective function (red, solid line) versus marker based reference data (blue,
dashed line). The x-direction is the walking direction.
B. Identification of objective weights
To take into account the different orders of magnitude of the
twelve optimality functions (21)–(32), we introduce a scaling
vector σ, which ensures that the elements of the weight vector
γ̃ := (γ T , pDSSS /σDSSS )T are of similar magnitude, to
enforce a well conditioned upper optimization problem. Here,
we set σ to
σ = {1, 1, 0.01, 10, 10, 1, 0.01, 0.1, 0.1, 0.001, 0.001, 0.1, 0.25},
and the initial weights to
γ̃ (0) = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.5, 1.0, 1.0, 1.0}.
For uniqueness reasons, we fix the first weight γ̃0 to one and
propose an initial ratio for double and single support of 1:4
(pDSSS = 0.25). Since the identification of pDSSS would be
infeasible for the case γ̃11 = 0 or γ̃12 = 0, we introduce the
bounds 0.1 ≤ γ̃11 ≤ 10, 0.01 ≤ γ̃12 ≤ 10. Furthermore, to
allow for both, minimization and maximization of the overall
single support duration, we define the bounds −10 ≤ γ̃9 ≤ 10.
For all other weighs we require non-negativity 0 ≤ γ̃i ≤ 10,
i = 0, . . . , 10, i 6= 9.
For the initial guess of γ̃ (0) we observe an average quadratic
deviation between motion capture data and computed optimal
Fig. 6. Phase durations resulting from identified objective function (red)
versus reference data (blue).
Phase durations, feet trajectories, and x-y coordinates of
the center of mass are satisfying reconstructed. The torso
orientation and the vertical CoM velocities (Figure 5, bottom
left) have a greater relative difference, which is due to their
small scaling and the fact that the upper level fitting is based on
absolute errors. This can be changed by adjusting the weight
vector ω, see Equ. (1).
The optimal motion for the identified objective function,
together with indicators of the reference motion, is given in
the figure sequence 7.
V. H UMANOID G AIT G ENERATION
Comparing the two humanoid robots to the human model
we can observe significant differences in architecture and
dynamical behavior.
The first robot, the iCub platform of Heidelberg University
(HeiCub), has 15 degrees of freedom and consists of an upper
body, a hip and two legs [15, 24]. It does neither have arms nor
Fig. 7. Optimal motion with respect to the identified weights. The motion of interest is shown by the 3D-SEA walker. The reference motion is
sketched by light colored lines as legs and a small ball as head. The head trajectory of the motion (which is not an explicit state) is only plotted
for the reference solution and is meant to simplify the comparison of the torso orientation. A corresponding video is available online: http://orb.iwr.uniheidelberg.de/ftp/CleverMombaur IOC RSS2016.
a head and is therefore quite similar to the considered template
model. It has a total height of 0.97m and a total mass of 26kg.
The second robot, HRP-2, is a full body bi-pedal robot with
30 degrees of freedom. It has a total height of 1.54m and a
total mass of 58kg [19, 18].
Most of the differences between the three embodiments
(human model, HeiCub (red), HRP-2 (gray)) can be directly
taken into account by adjusting the model parameters, model
constraints and environment constraints. Whereas we have
precise information about robot masses and limits, we have
no information about spring and damping constants in terms
of the considered SEA-model. In the following experiments
we use the same stiffness as for the human but reduce the
damping to 440 Ms/m. The relevant parameters and limits are
given in Table II.
TABLE II
PARAMETERS AND BOUNDS FOR ROBOT MODELS
body weight
est. upper body weight
foot weight
leg rest length
torso height
torso radius
spring stiffness
damping constant
bound ss duration (l,u)
bound ds duration (l,u)
safety margin to edge
(front, left, back, right)
HeiCub (red)
HRP-2 (gray)
26kg
12.5kg
0.8 kg
0.6m
0.38m
0.38m
3300kg/s2
440 Ms/m
0.5s, 1.5s
0.15s, 0.5s
(0.10, 0.05, . . .
. . . , 0.05, 0.05)m
58kg
28kg
1.2 kg
0.75m
0.8m
0.4m
3300kg/s2
440 Ms/m
0.5s, 1.5s
0.15s, 0.5s
(0.12, 0.08, . . .
. . . , 0.08, 0.08)m
We set up two different step stone scenarios, one for each
robot. Here, the second scenario is particularity interesting as it
has already been accomplished by HRP-2 14 at CNRS-LAAS,
using a standard pattern generator.
A. Transfer rules
Performing motion transfer based on optimality, as it is
done in this paper, it is desirable to transfer the identified
weights of human optimality criteria as direct as possible to
the humanoid objective function. However, as robots are much
more constrained than humans a direct transfer of objective
weights usually drives the robot model to some of its bounds.
Those active constraints can change the nature of the motion
dramatically. Therefore, we aim to identify transfer rules in
terms of weight re-scaling which prevent the model to be stuck
on a certain bound.
In our specific example (as most likely for many other ones)
the identified weight for the minimization of single support
phase duration is much too large for the robot and would
result in phase durations which are forced to its lower bound.
However, this would mean the same as predefining phase
durations and destroys the advantage of state dependent phase
transitions. Therefore, it is reasonable to not only adjust the
lower bounds for the phase duration, but also to decrease the
weight γ̃9 . Similar observations are made if we only include
stricter bounds on the swing foot velocity. Here, an active
bound results in an undesired M-shaped swing foot trajectory.
Therefore, in addition to the stricter bound the weight γ̃6
has to be increased. For the first robot model (HeiCub) the
weight γ̃9 is divided by two (γ̃9,rob1 = 0.25), which is
the only modification we introduce for this robot. For the
second robot model (HRP-2), where we define componentwise bounds on the CoM velocity of [−0.5, 0.5]m/s and
on the swing foot velocities of [−0.8, 0.8]m/s, we increase
γ̃6 to γ̃6,rob2 = 2 and set γ̃9 to the high negative value
of γ̃9,rob2 = −6. Furthermore, we include the fact that
in previous experiments (using a standard pattern generator)
HRP-2 successfully crossed the considered step stone scenario
with a ratio of double to single support of 1/7. Hence, for the
HRP-2 model we set the parameter pDSSS,rob2 = 0.14. Also
modifications in environment constraints might require a rescaling of optimality parameters. In our example, HRP-2 is
asked to manage steps with significantly greater differences in
step height than in the training step stone scenario. Because
the actuation of the SEA also mimics desired knee bending,
setting γ1 = 1.3 enforces a gait, where the robot model swings
its almost stretched legs outwards instead of swinging the foot
on a more direct line between step stones and adjusting the
leg length. To this end we divide this weight by two and set
γ1,rob2 = 0.65. In scenario 1 the topology of the step stones
is quite similar to the human example, except for the smaller
distances and the fact, that the robot has to start the motion
with the other leg. Due to the characteristics of the template
model this issue does not require any additional modifications,
neither in the dynamics of the model, nor in the constraints,
nor in the objective.
B. Gait synthesis
durations for the two robot models, see Figure 8. Exemplary
for all the quantities, the CoM trajectories of the human
motion and the humanoid motions are presented in Figure
9. Whereas the human motion is quite fast (v∅ = 0.43m/s),
the computed motion for the robots, with the reduced weight
for time minimization, results in much slower motions (both
v∅ = 0.18m/s), see Figure 9 and 10.
VI. C ONCLUSIONS
Fig. 8. Robot motions as result of an optimal control problem in terms
of CoM and feet trajectories, torso orientation and phase durations. Objective
weights, model parameters and bounds are adjusted to the constraints of
the two robots. A corresponding video is available online: http://orb.iwr.uniheidelberg.de/ftp/CleverMombaur IOC RSS2016.
We now use the same optimal control model which we have
used in the lower level of the inverse optimal control approach,
but adjust parameter and objective scaling as described in Table II and Subsection V-A. As we do not have a sophisticated
guess for an initial motion (as we had for the inverse optimal
control based on the motion capture experiments) we start with
a simple initial motion, where feet and CoM trajectories are
linearly interpolates between step postures and where the torso
orientation is constant to zero. Finally, we predefine the desired
values for the initial horizontal CoM position (xM , yM ) at
time t = 0 and the final ones at time T .
Fig. 9. CoM trajectories resulting from identified objective function for
the human model (blue) and the two robot models, HeiCub (red) and HRP-2
(black).
On basis of a step stone scenario, we have presented
an approach, which allows to transfer walking motions in
constrained environment from humans to humanoids that have
different embodiments and are moving within a new scenario
with significantly different properties.
Even though the optimal motion resulting from the identified optimization strategy does not exactly match the one
based on motion capture markers, the inverse optimal control
algorithm shows a satisfying performance. The average error
in trajectories and timing between the computed and the
reference motion is reduced by 40% in comparison to the
best hand tuned solution. Taking 121 upper level iterations for
full convergence, already after half of the iterations, suitable
weights are identified. More precisely, this means that for a
loss of accuracy by approximately 5% computational time can
be reduced by 50%. We have shown, that the optimal control
based transfer of motion leads to reasonable robot motions
for two different robots. We can observe the advantage of
physically interpretable optimization criteria to derive heuristic
(but physically motivated) transfer rules. An important step of
future research is a systematic investigation and formulation of
similar transfer rules which hold for a wide range of scenarios
and robots. Follow up work will focus on two issues. On
the computational side, we extend of our studies to different
walking scenarios and augment the optimality identification
across different trials and different subjects. On the practical
side, we investigate on how well the computed robot motions
transfer to their real counterparts. In this context, the computed
motions are not planned to be run open loop on the robot.
Rather they substitute existing methods (e.g. based on the table
cart model) to generate the input for the robot specific whole
body motion generation tools.
ACKNOWLEDGMENTS
Fig. 10. Phase durations resulting from identified objective function for
the human model (blue) and the two robot models, HeiCub (red) and HRP-2
(white).
The result of this computation are optimal trajectories for
the center of mass, the feet, the torso angles and the phase
The research leading to these results has received funding from the EU 7th Framework Program (FP7/2007-2013)
under grant agreement no 611909 (KoroiBot). We thank the
Simulation and Optimization group of H. G. Bock at the
University of Heidelberg for providing the optimal control
code Muscod-II. Furthermore we thank our KoroiBot partners:
M. Giese (CIN Tübingen) and team for performing the motion
capture experiments, T. Asfour (H2T, KIT) and team for
maintaining the motion capture database, and P. Souères and
O. Stasse (CNRS-LAAS) for their support towards the transfer
of concept to the real robot HRP-2.
R EFERENCES
[1] KoroiBot Motion Capture Database. https://koroibotmotion-database.humanoids.kit.edu/. Last visited April
2016.
[2] S. Albrecht, P. Basili, S. Glasauer, M. Leibold, and
M. Ulbrich. Modeling and analysis of human navigation
with crossing interferer using inverse optimal control. In
Proceedings of the Vienna International Conference on
Mathematical Modeling, 2012.
[3] R. M. Alexander. The gaits of bipedal and quadrupedal
animals. The International Journal of Robotics Research,
3(2):49–59, 1984.
[4] R. M. Alexander. Optima for animals. Princeton
University Press, 1996.
[5] G. B. Allende and G. Still. Solving bilevel programs with
the KKT-approach. Mathematical programming, 138(12):309–332, 2013.
[6] L. Boutin, A. Eon, S. Zeghloul, and P. Lacouture. From
human motion capture to humanoid locomotion imitation
application to the robots HRP-2 and HOAP-3. Robotica,
29(02):325–334, 2011.
[7] D. Clever and K. Mombaur. A new template model
for optimization studies of human walking on different
terrains. In Humanoid Robots (Humanoids), 2014 14th
IEEE-RAS International Conference on, pages 500–505.
IEEE, 2014.
[8] D. Clever and K. Mombaur. On the relevance of common
humanoid gait generation strategies in human locomotion
- an inverse optimal control approach. In Hoang X.P.
Rannacher R. Schlöder J. Bock, H.G., editor, Modeling,
Simulation and Optimization of Complex Processes HPSC 2015 (to appear). Springer, 2016.
[9] D. Clever, K. Hatz, and K. Mombaur. Studying Dynamical Principles of Human Locomotion using Inverse
Optimal Control. In PAMM, volume 14, pages 801–802.
Wiley Online Library, 2014.
[10] P. De Leva. Adjustments to Zatsiorsky-Seluyanov’s
segment inertia parameters. Journal of biomechanics, 29
(9):1223–1230, 1996.
[11] D. Gonzalez-Aguirre, M. Vollert, T. Asfour, and R. Dillmann. Robust Real-time 6D Active-visual Localization
for Humanoid Robots. In IEEE International Conference
on Robotics and Automation (ICRA 2014), pages 2785–
2791, 2014.
[12] K. Hashimoto, H. Kang, M. Nakamura, E. Falotico,
H. Lim, A. Takanishi, C. Laschi, P. Dario, and
A. Berthoz. Realization of biped walking on soft ground
with stabilization control based on gait analysis. In
Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ
International Conference on, pages 2064–2069. IEEE,
2012.
[13] K. Hatz. Efficient numerical methods for hierarchical
dynamic optimization with application to cerebral palsy
gait modeling. PhD thesis, University Heidelberg, 2014.
[14] K. Hatz, J. P. Schlöder, and H. G. Bock. Estimating
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
Parameters in Optimal Control Problems. SIAM Journal
on Scientific Computing, 34(3):A1707–A1728, 2012.
Y. Hu, K. Mombaur, and F. Nori. Using optimal control
to generate squat motions for the humanoid robot iCub
with SEA. Proceedings of Dynamic Walking, 2015.
S. G. Johnson. The NLopt nonlinear-optimization package. http://ab-initio.mit.edu/nlopt, 2014.
S. Kajita, F. Kanehiro, K. Kaneko, K. Fujiwara, K. Yokoi,
and H. Hirukawa. A realtime pattern generator for
biped walking. In Robotics and Automation, 2002.
Proceedings. ICRA’02. IEEE International Conference
on, volume 1, pages 31–37. IEEE, 2002.
S. Kajita, T. Nagasaki, K. Kaneko, K. Yokoi, and
K. Tanie. A running controller of humanoid biped HRP2LR. 2005.
K. Kaneko, F. Kanehiro, S. Kajita, K. Yokoyama,
K. Akachi, T. Kawasaki, S. Ota, and T. Isozumi. Design
of prototype humanoid robotics platform for HRP. In
Intelligent Robots and Systems, 2002. IEEE/RSJ International Conference on, volume 3, pages 2431–2436. IEEE,
2002.
A Koenemann, A. Del Prete, Y. Tassa, E. Todorov,
O. Stasse, M. Bennewitz, and N. Mansard. Experiments
with MuJoCo on HRP-2. In IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS
2015), 2015.
M. Kudruss, M. Naveau, O. Stasse, N. Mansard,
C. Kirches, P. Souères, and K. Mombaur. Optimal Control for Whole-Body Motion Generation using Centerof-Mass Dynamics for Predefined Multi-Contact Configurations. In IEEE/RAS International Conference on
Humanoid Robots (Humanoids 2015), page 7, November
2015.
C. Mandery, Ö. Terlemez, N. Vahrenkamp, and T. Asfour.
The Koroibot Markerset. Technical report, Institute for
Anthropomatics and Robotics, Karlsruhe Institute for
Technology (KIT), Karlsruhe, Germany, 2014.
C. Mandery, Ö. Terlemez, M. Do, N. Vahrenkamp,
and T. Asfour. The KIT Whole-Body Human Motion
Database. In IEEE International Conference on Advanced Robotics (ICAR 2015), pages 329–336, 2015.
G. Metta, L. Natale, F. Nori, G. Sandini, D. Vernon,
L. Fadiga, C. Von Hofsten, K. Rosander, M. Lopes,
J. Santos-Victor, et al. The iCub humanoid robot: An
open-systems platform for research in cognitive development. Neural Networks, 23(8):1125–1134, 2010.
K. Mombaur, H. G. Bock, and J. P. Schlöder. Open-loop
stable solutions of periodic optimal control problems in
robotics. ZAMM - Journal of Applied Mathematics and
Mechanics, 85(7):499–515, 2005.
K. Mombaur, A. Truong, and J.-P. Laumond. From
Human to Humanoid Locomotion An inverse optimal
control approach. Autonomous Robots, 28(3), 2010.
F. Nori, N. Kuppuswamy, and S. Traversaro. Simultaneous state and dynamics estimation in articulated
structures. In IEEE/RSJ International Conference on
[28]
[29]
[30]
[31]
[32]
Intelligent Robots and Systems (IROS 2015), pages 3380–
3386, 2015.
A. Pajon, G. De Magistris, S. Miossec, K. Kaneko, and
A. Kheddar. A humanoid walking pattern generator
for sole design optimization. In IEEE International
Conference on Advanced Robotics (ICAR 2015), pages
105–110, 2015.
T. Park and S. Levine. Inverse optimal control for
humanoid locomotion. In Robotics Science and SystemsWorkshop on Inverse Optimal Control and Robotic
Learning from Demonstration, 2013.
M. J. D. Powell. The BOBYQA algorithm for bound
constrained optimization without derivatives. 2009.
O. Ramos, M. Garcia, N. Mansard, O. Stasse, J. B. Hayet,
and P. Souères. Towards reactive vision-guided walking
on rough terrain: an inverse dynamic based approach.
International Journal of Humanoid Robotics (IJHR), 11:
1004–1026, 2014.
M. Sreenivasa, K. Mombaur, and J. P. Laumond. Walking
paths to and from a goal differ: On the role of bearing
angle in the formation of human locomotion paths. Plos
one, 10(4), 2015.